Predicting risk-adjusted mortality for trauma patients: Logistic versus multilevel logistic models

Background: We sought to determine whether random-intercept multilevel (ML) regression modeling might be useful in the assessment of hospital care for trauma patients. Theoretical advantages of ML over standard logistic regression (LR) include separating variability due to patient-level and hospital-level predictors, “shrinkage” of estimates for lower-volume hospitals toward the overall mean, and fewer hospitals falsely identified as outliers.

Methods: We used Nationwide Inpatient Sample (NIS) data from 2002-4 to construct LR models of hospital mortality after admission with a principal ICD-9-CM injury diagnosis (ICD-9-CM 800-904, 910-929, 940-957, 959). After considering various predictors, we used patient-level indicator variables for age groups, sex, maximum AIS for the head region (3,4,5), maximum AIS for other body regions (3,4,5), and mechanisms (fall, gunshot, motor vehicle). Using standard LR and MLLR, we compared predictions based upon 2002, 2003, and 2004 data to actual mortality observed in the same hospitals in 2004, 2005, and 2006 NIS respectively.

Results: Patient-level fixed effects were similar for either method in all years, with mortality associated most strongly with AIS=5 head injury, other AIS=5 injury, or higher age groups. ML models identified fewer hospitals as outliers. Differences between actual and predicted mortality were significantly smaller with MLLR models compared to standard LR models.

Conclusions: ML models may have advantages for the measurement and explanation of interhospital differences in trauma patient outcomes.